Image Classification on ImageNet-1k (val) (Top-1 Accuracy and No-Crop Resize Accuracy)
90.7Top-1 Acce/14 ema
Evaluation results
| Method | Links | ||
|---|---|---|---|
| e/14 emaModel backbone=e/14, Fine-tuned resolution=560px, Polyak averaging (EMA)=true, Training setup=Fine-tuned on ImageNet2023.02 | 90.7 | 90.84 | |
| 22B emaModel backbone=22B, Fine-tuned resolution=560px, Polyak averaging (EMA)=true, Training setup=Fine-tuned on ImageNet2023.02 | 90.62 | — | |
| 22BModel backbone=22B, Fine-tuned resolution=560px, Polyak averaging (EMA)=false, Training setup=Fine-tuned on ImageNet2023.02 | 90.6 | — | |
| 22BModel backbone=22B, Fine-tuned resolution=384px, Polyak averaging (EMA)=false, Training setup=Fine-tuned on ImageNet2023.02 | 90.44 | — | |
| e/14 emaModel backbone=e/14, Fine-tuned resolution=384px, Polyak averaging (EMA)=true, Training setup=Fine-tuned on ImageNet2023.02 | 90.44 | — | |
| G/14 emaModel backbone=G/14, Fine-tuned resolution=518px, Polyak averaging (EMA)=true, Training setup=Fine-tuned on ImageNet2023.02 | 90.33 | 90.47 | |
| g/14 emaModel backbone=g/14, Fine-tuned resolution=518px, Polyak averaging (EMA)=true, Training setup=Fine-tuned on ImageNet2023.02 | 90.25 | 90.11 | |
| L/16Model backbone=L/16, Fine-tuned resolution=384px, Polyak averaging (EMA)=false, Training setup=Fine-tuned on ImageNet2023.02 | 88.6 | — | |
| B/16Model backbone=B/16, Fine-tuned resolution=384px, Polyak averaging (EMA)=false, Training setup=Fine-tuned on ImageNet2023.02 | 87.02 | — | |
| 22BModel backbone=22B, Training setup=JFT-only (zero-shot)2023.02 | 78.5 | — | |
| FP32Backbone=SigLIP2, Quantization Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 78.47 | — | |
| RepQ-ViT (RepQ) + RegCacheBackbone=SigLIP2, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 77.13 | — | |
| e/14Model backbone=e/14, Training setup=JFT-only (zero-shot)2023.02 | 76.7 | — | |
| G/14Model backbone=G/14, Training setup=JFT-only (zero-shot)2023.02 | 76.6 | — | |
| RepQ-ViT (RepQ)Backbone=SigLIP2, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 76.43 | — | |
| g/14Model backbone=g/14, Training setup=JFT-only (zero-shot)2023.02 | 76.3 | — | |
| FP32Backbone=SigLIP, Quantization Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 76.05 | — | |
| RepQ-ViT (RepQ) + RegCacheBackbone=SigLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 75.94 | — | |
| RepQ-ViT (RepQ)Backbone=SigLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 75.23 | — | |
| SmoothQuant (SQ) + RegCacheBackbone=SigLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 74.65 | — | |
| L/16Model backbone=L/16, Training setup=JFT-only (zero-shot)2023.02 | 73.9 | — | |
| SmoothQuant (SQ) + RegCacheBackbone=SigLIP2, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 72.45 | — | |
| B/16Model backbone=B/16, Training setup=JFT-only (zero-shot)2023.02 | 70.7 | — | |
| SmoothQuant (SQ)Backbone=SigLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 70.55 | — | |
| FP32Backbone=OpenCLIP, Quantization Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 70.22 | — | |
| RepQ-ViT (RepQ) + RegCacheBackbone=OpenCLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 70.06 | — | |
| NaïveBackbone=SigLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 69.71 | — | |
| RepQ-ViT (RepQ)Backbone=OpenCLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 68.7 | — | |
| FP32Backbone=CLIP, Quantization Precision=FP32, Evaluation Protocol=Zero-shot2025.10 | 68.32 | — | |
| RepQ-ViT (RepQ) + RegCacheBackbone=CLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 68.1 | — | |
| RepQ-ViT (RepQ)Backbone=CLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 67.39 | — | |
| SmoothQuant (SQ) + RegCacheBackbone=OpenCLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 67.07 | — | |
| SmoothQuant (SQ) + RegCacheBackbone=CLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 60.3 | — | |
| SmoothQuant (SQ)Backbone=OpenCLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 46.5 | — | |
| NaïveBackbone=OpenCLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 46.12 | — | |
| SmoothQuant (SQ)Backbone=CLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 35.54 | — | |
| NaïveBackbone=CLIP, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 34.01 | — | |
| SmoothQuant (SQ)Backbone=SigLIP2, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 28.2 | — | |
| NaïveBackbone=SigLIP2, Quantization Precision=W8A8, Evaluation Protocol=Zero-shot2025.10 | 26.04 | — |